From f4170f78bdfe69050c60bf3b51370d6c97593489 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Fri, 20 Dec 2024 09:12:48 +0100 Subject: IQ3_XXS_R4 (#153) * iq3_xxs_r4: 1st shot on Zen4 PP-512: 107 t/s -> 137 t/s TG-128(1 thread): 2.64 t/s -> 3.44 t/s * iq4_xxs_r4: WIP * iq4_xxs_r4: 1st shot at AVX2 Note: there is a bug in the AVX2 implementation for nrc_y = 1 for IQ quants with blocks of 32. I have fixed it for now by using the nrc_y > 1 implementation (which works) also for nrc_y = 1. * iq3_xxs_r4: NEON --------- Co-authored-by: Iwan Kawrakow --- ggml/include/ggml.h | 2 ++ 1 file changed, 2 insertions(+) (limited to 'ggml/include/ggml.h') diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h index 1a4e7f19..90d34b40 100644 --- a/ggml/include/ggml.h +++ b/ggml/include/ggml.h @@ -418,6 +418,7 @@ extern "C" { GGML_TYPE_Q4_K_R4 = 212, GGML_TYPE_Q5_K_R4 = 213, GGML_TYPE_Q6_K_R4 = 214, + GGML_TYPE_IQ3_XXS_R4= 218, GGML_TYPE_IQ4_NL_R4 = 220, GGML_TYPE_IQ4_XS_R4 = 223, GGML_TYPE_BF16_R16 = 230, @@ -496,6 +497,7 @@ extern "C" { GGML_FTYPE_MOSTLY_Q4_K_R4 = 212, // except 1d tensors GGML_FTYPE_MOSTLY_Q5_K_R4 = 215, // except 1d tensors GGML_FTYPE_MOSTLY_Q6_K_R4 = 214, // except 1d tensors + GGML_FTYPE_MOSTLY_IQ3_XXS_R4= 217, // except 1d tensors GGML_FTYPE_MOSTLY_IQ4_NL_R4 = 219, // except 1d tensors GGML_FTYPE_MOSTLY_IQ4_XS_R4 = 222, // except 1d tensors GGML_FTYPE_MOSTLY_BF16_R16 = 224, // except 1d tensors -- cgit v1.2.3